Detecting and Classifying Humanitarian Crisis in Arabic Tweets

摘要

Yemen and Syria are suffering from the worst humanitarian crisis in the world. Since 2016, 80% of the population in Yemen are dying from hunger, and 3,886 died from cholera. While since 2011, 65% of the Syrian population have become refugees. During these crises, people from both countries turned to Twitter to convey their crisis-related messages. Humanitarian organizations have realized the effectiveness of gathering, analyzing, and classifying tweets' contents to enhance their crisis rescue plan. However, most of the available crisis resources are either in the English language or cover hazards and natural disasters only. Also, there is a lack of knowledge of the most common terms used for crisis description by Arabic users. So, organizations found it difficult to gather, annotate, preprocess, extract features, and classifying Arabic crisis tweets content. As a result, there is a delay in responding to famine, cholera, and refugee crisis and a lot of loss in lives. The paper aims to proposed methodologies for extracting unique crisis terms, building annotation criteria, and enhancing classification for crisis-related messages in the Arabic language. Also, we produced a humanity crisis corpus for classifying tweets in Arabic. For that, we used keywords from each topic produced by the LDA model to collect crisis tweets. Then, we built crisis annotation criteria guided by a unique word list generated from word embedding models. Finally, we combined features from topics, words, and sentences then implemented by supervised methods for classification. Results indicate that our proposed methods enhance the classification model's performance. Besides, it increases the classifier's ability to detect more positive crisis classes to the right label. On the other hand, this paper provides humanitarian organizations with tools and methods for Arabic crisis-messages classification in social media and opens new opportunities for future studies in crisis management.